Question Raptor Lake - Official Thread

Page 137 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Hulk

Diamond Member
Oct 9, 1999
4,525
2,518
136
Since we already have the first Raptor Lake leak I'm thinking it should have it's own thread.
What do we know so far?
From Anandtech's Intel Process Roadmap articles from July:

Built on Intel 7 with upgraded FinFET
10-15% PPW (performance-per-watt)
Last non-tiled consumer CPU as Meteor Lake will be tiled

I'm guessing this will be a minor update to ADL with just a few microarchitecture changes to the cores. The larger change will be the new process refinement allowing 8+16 at the top of the stack.

Will it work with current z690 motherboards? If yes then that could be a major selling point for people to move to ADL rather than wait.
 
  • Like
Reactions: vstar

Hulk

Diamond Member
Oct 9, 1999
4,525
2,518
136
I mostly agree.

In the desktop, Intel did a great job of obtaining parity and even eclipsing Zen 4 performance in many cases by the use of "Big/Little" and by throwing power budget out the window. These trade-offs make sense in the desktop.

In the laptop market, my feeling is that Zen 4 and RDNA2 on TSMC N5 will be a much more trouble to Intel's Raptor Lake design (time will tell). Considering the market makeup of laptop vs desktop, this will likely be the more important battle between the 2 companies.

In the server market, AMD's design and scaling appear to dominate anything that Intel can put on the table throughout 2023. I suspect AMD would be perfectly willing to sacrifice market share in the PC Desktop to Intel in exchange for even more server market share.

For me, this comes down to Intel being able to meet its process roadmap. On paper, Intel is poised to dominate transistor density by 2024. My personal engineering opinion is that Intel's process roadmap is dangerously optimistic. Every process step involves big changes and advancements. Every process step is scheduled to coincide with a new CPU architecture. While this looks good on paper, will Intel be able to pull this off? Intel's transition to 10nm was a disaster. They are in the process of laying off a ton of engineers. To me, it seems like an impossibly difficult task that Intel executives are asking of Engineering. It took Intel 10 years to dig the hole they are in. I find it very hard to see a way for them to climb out of it in 3.

FWIW, I could easily be wrong. When Intel nose dived with Itanium and P4 I didn't see the Israel team bailing the company out with Conroe either ;). Of course, Intel still had the process advantage then.

from here:

https://www.granitefirm.com/blog/us/2021/12/28/tsmc-process-roadmap/

I agree completely. If you remember the move from 22nm to 14nm was even semi-disaster for Intel and I think one of the reasons why Broadwell was really just a mobile release.

10 years ago I would never have suspected Intel would not have a significant process advantage. Now, like you, I don't have faith they can execute on their node plan. We shall see.

As for mobile I find it strange that there are so few reviews of Alder Lake U parts. It's really hard to find benchmarks and even harder to find head-to-head comparisons of Intel vs. AMD current mobile parts.
 

jpiniero

Lifer
Oct 1, 2010
15,223
5,768
136

Hulk

Diamond Member
Oct 9, 1999
4,525
2,518
136
How do you find your regular workflow with the 13900K? All complaints regarding E-core performance gone?

I love the 13900K. It's a beast. I have it set at 175W and on lightly threaded workloads 2 P's will ramp up to 5.8GHz. It's also great for multitasking. For example, I'll be frameserving a Vegas Pro timeline to Handbrake for rendering using those 16E's while doing Photoshop work with the P's in the foreground. If I do a compute heavy task like perspective correction I'll hear the CPU fan ramp up for like 2 seconds and then shut down. All the while my editing experience is seamless and by that I mean real time, I'm not waiting on the computer, which makes for a more creative and fun edit.

As for "TSW." I have an old pair of AR TSW610 from college. I still love them. The joke from one of the engineers supposedly was that TSW stood for "This Sh*t Works!" In reality the advertising said TitaniumSolidWood, which makes more sense of course.

Anyway with 16E's available TSW!
 

Hulk

Diamond Member
Oct 9, 1999
4,525
2,518
136
I have to say for the PC enthusiast the hours of testing alone is worth the price of admission for my 13900K. I've finally managed to "corner" those E's! Don't know why I didn't think of this before.

Here's the setup. 1+16 with the E's running 4.3 and the P running 1.4GHz, it's basically doing nothing and consuming about 2W.

First fun fact. At 4.3GHz an E core uses about 7.5 Watts. I'm fairly confident about that number.

Second. Running CPUmark99 the Thread Director correctly runs it on an E core as the result is 727, which is solidly in Skylake range. A P at 1.4 would give a much lower result. Also, per GHz Gracemont has better IPC in CPUmark99 than Golden Cove or Raptor Cove.

Third. Cinebench ST for an E at 4.3 is 1171. Finally a number I can believe. Again, solidly in Skylake territory.

Finally the Handbrake test comes in at 224 seconds. Given that the one P core at 1.4GHz isn't doing much these 16E's at 4.3 GHz are faster than a 11900K (219 seconds), 10900K (234 seconds), and 5800X (240 seconds). My old 4770K ran this test in 813 seconds. So the E cores in the 13900K are like tacking on 3.5 4770K's to the Golden Coves.
 

nicalandia

Diamond Member
Jan 10, 2019
3,331
5,282
136
If the P's and E's are fabricated on the same process how have they increased the E to P ratio from the current 3.1:1 to 4:1? I assume either the P's are getting larger or the E's are getting smaller or a combination of both?

Different part of the core scale differently since The P core and E cores have different design they will scale differently, but nothing as drastic as you believe, The Golden Cove to Gracemont Core ratio is still about 4:1(7.123 mm^2 vs 1.5 mm^2 so technically 4.5:1), but those little e cores are not all by themselves are they right?

They are in a Cluster(Quad Core Cluster to be precise), the cluster size is 8.296 mm^2 wich is 16% larger than a single Golden Cove.

Now to the Meteor Lake side: Redwood Cove size is 5.33 mm^2 and The Crestmont core size is 1.046 mm^2 which is 5:1, but as mentioned before those little e cores come in clusters, so for MTL the cluster size is 5.907 which is 11%


I would like to point out that the only source we have on Redwood Cove/Crestmont comes from a mobile die which is likely MTL-H, which restricts Crestmont cores to 3MiB of L2, I suspect that MTL-S Crestomont cores to have either 6MiB of L2 per Cluster(or at the very least 4MiB), that will increase the cluster size past 6 mm^2 bringing the ration back to what we currently are with Raptor Cove/Raptormont


1667766404839.png
Source: https://semianalysis.substack.com/p/meteor-lake-die-shot-and-architecture
 
  • Like
Reactions: Tlh97 and coercitiv

JayMX

Member
Oct 18, 2022
31
73
51

dullard

Elite Member
May 21, 2001
25,510
4,004
126
They used the exact same RAM for Raptor Lake, which we also know benefits from faster memory.

And they also used the exact same cooler, which by your insistence would be "crippling" Raptor Lake more.

But even the slightest win for Intel has to be a conspiracy in your mind...
I came to post this. Puget themselves say for photoshop that DDR5 memory speeds have a small difference. But of course, that means they are biased against AMD when both systems have memory lower than their fastest supported memory. :rolleyes:
1667928849040.png
Interestingly enough, DDR5 doesn't make all that much of a difference in Photoshop. In most of the other applications we tested, DDR5 gave roughly a 10% increase in performance over DDR4, but for Photoshop the difference was within the margin of error.
1667929130937.png
 

JayMX

Member
Oct 18, 2022
31
73
51
They used the exact same RAM for Raptor Lake, which we also know benefits from faster memory.

If that holds true how do you explain the following numbers:

- Puget
13900k(DDR5 4800): 1660 overall score
7950x (DDR5 4800): 1532 overall score
diff: 8%
AP_PS.png

- The Verge
13900k(DDR5 6600): 1517 overall score
7950x (DDR5 6000): 1497 overall score
diff: 1%
(https://www.theverge.com/23410428/intel-core-i9-13900k-review)

- Techspot
13900k(DDR5 6400): 1612 overall score
7950x (DDR5 6000): 1523 overall score
diff: 6%
(https://www.techspot.com/review/2552-intel-core-i9-13900k/)

Both The Verge and Techspot used 360 AIO for cooling. All three testers run "PugetBench for Photoshop".

So how can a "decent cooling/ slow DDR5" system outperform a "good cooling/ fast DDR5 " system?
(Puget system got the top score)
 
  • Like
Reactions: Tlh97 and Markfw

TheELF

Diamond Member
Dec 22, 2012
4,027
753
126
If that holds true how do you explain the following numbers:

- Puget
13900k(DDR5 4800): 1660 overall score
7950x (DDR5 4800): 1532 overall score
diff: 8%
View attachment 70610

- The Verge
13900k(DDR5 6600): 1517 overall score
7950x (DDR5 6000): 1497 overall score
diff: 1%
(https://www.theverge.com/23410428/intel-core-i9-13900k-review)

- Techspot
13900k(DDR5 6400): 1612 overall score
7950x (DDR5 6000): 1523 overall score
diff: 6%
(https://www.techspot.com/review/2552-intel-core-i9-13900k/)

Both The Verge and Techspot used 360 AIO for cooling. All three testers run "PugetBench for Photoshop".

So how can a "decent cooling/ slow DDR5" system outperform a "good cooling/ fast DDR5 " system?
(Puget system got the top score)
Puget actually uses intel settings while the others use "default out-of-the-box" they get thermal throttle and drop performance, while the puget system runs cooler and thus boosts higher all by itself.
They do the same for AMD systems.

They also explain why they use the ram they use.

Second, we are using DDR5-4800MHz memory for both the Intel 12th/13th Gen processors, as well as Ryzen 7000. While 4800MHz is the fastest supported RAM speed for the 12th Gen, the new 13th Gen processors technically support up to DDR5-5600MHz if you have one, single rank stick per channel, or DDR5-5200MHz otherwise.
Similarly, Ryzen 7000 supports DDR5-5200MHz if using one stick per channel, but without overclocking is technically limited to just DDR5-3600MHz if using 2 sticks per channel.

This mess of what RAM speeds are officially supported based on how many sticks you are using, and whether they are single or dual rank, makes it extremely hard to standardize on a single configuration. Luckily, most of the applications we test are not significantly affected by RAM speed as long as you don't take it too far, so we decided on using DDR5-4800MHz across the board.
 

Abwx

Lifer
Apr 2, 2011
11,557
4,349
136
So speaking within the bounds of stock, dual channel configs, they're actually speccing the Intel system below its max, while overclocking the AMD system. This is what some people here call "handicap[ping] the AMD chip", not that I'd expect a shred of intellectual honest from them at this point.

The bound is elsewhare than in the config settings, basically they use a software where Intel perform well, wich is not a problem, but then set it such that it doesnt use more than 8 threads, this way a 5950X FI cant make full usage of it s better scaling over 8T.
 

Exist50

Platinum Member
Aug 18, 2016
2,452
3,102
136
The bound is elsewhare than in the config settings, basically they use a software where Intel perform well, wich is not a problem, but then set it such that it doesnt use more than 8 threads, this way a 5950X FI cant make full usage of it s better scaling over 8T.
Where do you see any such limitation on thread count?
 

Abwx

Lifer
Apr 2, 2011
11,557
4,349
136
Where do you see any such limitation on thread count?

Because the 7700X perform the same as a 7950X.

Count ST perf as 100 and SMT gain as 20%, the 7700X will output 800 on 8 threads, if there s 10 threads it will output 840.

On the other hand a 7950X output 1000 on 10 threads, and there s no such delta between those two CPUs, when looking at the 7600X score it s clearly visiible that it s even barely 8 threads.
 

Exist50

Platinum Member
Aug 18, 2016
2,452
3,102
136
Because the 7700X perform the same as a 7950X.

Count ST perf as 100 and SMT gain as 20%, the 7700X will output 800 on 8 threads, if there s 10 threads it will output 840.

On the other hand a 7950X output 1000 on 10 threads, and there s no such delta between those two CPUs, when looking at the 7600X score it s clearly visibible that it s even barely 8 threads.
So it's not that they impose some arbitrary limit on thread count, but rather that the applications in question are simply not embarrassingly parallel. That in no way invalidates the results.
 

Abwx

Lifer
Apr 2, 2011
11,557
4,349
136
So it's not that they impose some arbitrary limit on thread count, but rather that the applications in question are simply not embarrassingly parallel. That in no way invalidates the results.

It s not that there s only a low thread count, it s that for some reason Photoshop seems to be exclusively optimised for Intel, FI a 12600K roughly match a 7950X in this app.

Computerbase tests with Photoshop show that there s something that is cache dependant since a 5800X3D perform better than the 10% higher clocked 5950X, rest is specific optimisation for Intel to explain such a discrepancy.

The difference between the 13900K and the 7950X is at least 60% better for the former, wich is telling that Photoshop is all but an equally optimised soft for both uarch, actually that s optimisation for one and litteraly unoptimisation for the other...

 

TheELF

Diamond Member
Dec 22, 2012
4,027
753
126
Because the 7700X perform the same as a 7950X.

Count ST perf as 100 and SMT gain as 20%, the 7700X will output 800 on 8 threads, if there s 10 threads it will output 840.

On the other hand a 7950X output 1000 on 10 threads, and there s no such delta between those two CPUs, when looking at the 7600X score it s clearly visiible that it s even barely 8 threads.
You can't take a single thread score, multiply it, and come to any sort of conclusion.
It's not just the clocks that changes depending on how many cores are doing full work, or the SMT factor that changes depending on the workload, or how many different threads do different work.
It's also the cache thrashing and the mem bandwidth that different software create/need, let alone the cross CCX lag that can happen.

CPUs with fewer cores do run higher clocks and are often faster, we see that in game benches all the time.
 

Abwx

Lifer
Apr 2, 2011
11,557
4,349
136
You can't take a single thread score, multiply it, and come to any sort of conclusion.
It's not just the clocks that changes depending on how many cores are doing full work, or the SMT factor that changes depending on the workload, or how many different threads do different work.
It's also the cache thrashing and the mem bandwidth that different software create/need, let alone the cross CCX lag that can happen.

CPUs with fewer cores do run higher clocks and are often faster, we see that in game benches all the time.


Same tests from Computerbase show the 8C 7700X at 1-2% better IPC on MT than a 8 + 0 12900K/13700K, so 70% advantage for a 13900K vs a 7950X is nothing that has to do with the CPU perf but with the software being exclusively, and highly, optimised for one uarch.
 
Jul 27, 2020
20,040
13,738
146
nothing that has to do with the CPU perf but with the software being exclusively, and highly, optimised for one uarch.
That's nothing new. Intel devotes a lot of resources to ensure that software is well optimized for their CPUs, even from third party vendors of popular software. That's the benefit they get from having the lion's share of the market (not talking about enthusiasts but in general, office PCs and typical households with no little to no IT knowledge).

Software developers will most likely be given an Intel corporate PC and that's what they will write their software on and try to optimize for. I don't suppose there are many software publishers that go to great lengths to ensure that their software runs well on AMD CPUs. It's also not bad in a sense. It forces AMD engineers to design their architecture based on the typical instruction mix of existing software rather than relying on developers to optimize for their CPUs which may or may not happen. I think AMD is doing really well performance wise (especially in AVX-512) for a company that a large percentage of the public may not even be aware of.
 

TheELF

Diamond Member
Dec 22, 2012
4,027
753
126
It s not that there s only a low thread count, it s that for some reason Photoshop seems to be exclusively optimised for Intel, FI a 12600K roughly match a 7950X in this app.

Computerbase tests with Photoshop show that there s something that is cache dependant since a 5800X3D perform better than the 10% higher clocked 5950X, rest is specific optimisation for Intel to explain such a discrepancy.

The difference between the 13900K and the 7950X is at least 60% better for the former, wich is telling that Photoshop is all but an equally optimised soft for both uarch, actually that s optimisation for one and litteraly unoptimisation for the other...

Same tests from Computerbase show the 8C 7700X at 1-2% better IPC on MT than a 8 + 0 12900K/13700K, so 70% advantage for a 13900K vs a 7950X is nothing that has to do with the CPU perf but with the software being exclusively, and highly, optimised for one uarch.
What numbers are you referring to?!
Because the computerbase bench you link to doesn't even include photoshop, it does have photoscan but the numbers don't match what you are saying.

The numbers do match to adobe premiere, same company different app.
But it's not 60% better for intel, it's 37% lower time for intel, or it does the task in ~60% of the time.
You always use the bigger number as the base for figuring out percentages.